<Thumos experiment> the dropout layer in train.prototxt was cancelled ,but i found overfitting when trying to test my caffemodel by using eval_detection.py

[fanxx17]

the dropout layer in train.prototxt was cancelled ,but i found overfitting when trying to test my caffemodel by using eval_detection.py:
however,the caffemodel provided in Thumos's README.md has beautiful performance. Does train.prototxt really nead dropout layer when you train Thumos? thank you for your reply
...
}
#layer {

name: "drop6"

type: "Dropout"

bottom: "fc6"

top: "fc6"

dropout_param {

dropout_ratio: 0.5

}

#}
...